Independence

Standing

Not affiliated with, endorsed by, or controlled by any company it measures

Protocol STAMP is published by Budday Budderson Studio LLC, a New Mexico company. It is not affiliated with, endorsed by, or controlled by any model vendor, serving provider or gateway named in its reports. No company has advance sight of results, right of reply before publication, or influence over what is measured. The Log reports what was actually charged, never an estimate.

It does accept support from companies it measures, under the conditions below. That is a change from an earlier version of this page, which said it accepted nothing from them at all; the change is recorded in Corrections + changes with its date and its reason. Current support from companies this program measures: none. That line is kept current. If it ever reads otherwise it will say so before a single credit is spent, not after.

No company may require approval of anything, attach a non-disclosure term to its support, or ask for a figure to be withdrawn. These are not preferences. They are terms every supporter agrees to in writing before anything is accepted.

Recusal

Rules that are code, not intentions

• A judge never grades its own family. The runner refuses before the first call and seats the reserve.
• A candidate never grades. A model being measured has an interest in the measurement.
• An anchor is not a judge. A model measured twice for two purposes is disclosed, not excluded.
• Graders are qualified by probe, not reputation, against the cases that have broken graders before.

Money

What support may look like, and what it may not

Cash is preferred, because cash is the only form that buys a model from a company that gives nothing. Support is also accepted as API credits, metered keys with a spending ceiling, cloud compute and hosting, services in kind, and reader contributions. Advertising, where it appears, is placed by a third party without knowledge of content and is confined to the pages listed in the privacy statement; it never appears on an issue. Support of this publication is not tax-deductible: the recipient is a for-profit company.

The ten conditions. These were written on Sep 07, 2026 and published here on Sep 08, 2026, before any application had been filed and while no company had offered anything. The date is the point of them. Rules written cold are principles. The same words written the week a laboratory offers twenty thousand dollars are terms shaped to fit an offer, and no amount of careful drafting repairs that difference.

1. Everyone is asked, identically. Every company on the measurement list is asked on the same terms, in the same words, in the same week. This publication does not choose which company to approach, negotiate with, or be close to. The full request list is published before the asking, with the answers blank, so that it cannot be assembled afterwards to flatter anyone. Nothing has been asked of anyone yet; the table appears on this page before the first letter goes out.
2. A company’s support never measures that company. Credits or keys from a measured company are spent measuring other companies’ models, or on development work — probe drafting, ground-truth verification, re-grades. They are never spent on a published measurement of the giver. A company’s own models are always bought at retail with this publication’s own money. A laboratory may fund this publication’s view of its competitors. No laboratory funds its view of itself.
3. Self-measurement is bought blind. Where a provider’s terms allow it, a measured company’s models are purchased through a gateway on an account not identifiable as this publication’s, so that evaluation traffic cannot be recognized and handled differently from ordinary customer traffic. Where terms do not allow it, the issue says so.
4. Cash outweighs credits, and a reserve is held. Support from sources other than measured companies exceeds support from measured companies at every point in the year, and a cash reserve covering at least two months of the published cadence is maintained. A publication that cannot afford to lose a funder has already been bought. The reserve is what makes “publish it anyway” a decision rather than a phrase.
5. Concentration limits — proportions, never a ceiling. There is no upper limit on what this publication may raise, and no limit on how many supporters it may have. What is limited is concentration: no single measured company above 10% of the year’s total resources; all measured companies together at most 25%; any philanthropic funder at most 50% in the first year and 25% after; anonymous support at most 15%. Total resources means everything — cash, credits at list value, services in kind — not cash alone. Because these are ratios, raising more raises every funder’s permitted share in step: the rule constrains dependence, not ambition.
6. Credits are accepted only on private endpoints. Paid-tier or zero-data-retention endpoints only. Never a tier that logs, reviews, or trains on what is sent to it. The held-out probes are the instrument; an instrument that leaks is not one.
7. The split route. Where a credited route is used at all, one of a candidate’s two halves runs on the credited route and the other at retail. The halves are compared as they always are, and a divergence beyond the usual bounds is published as a finding rather than reconciled away.
8. Disclosure per issue, not per year. Every issue names, model by model, which route each measurement ran on and whose credit it was. A reader checking one figure can see how that figure was paid for without reconstructing it from an annual summary.
9. The objection clause, and what is not an objection. A correction is an assertion of fact with evidence behind it — this figure is wrong, here is the log. It is welcome from anyone, supporter or not, it is checked on the merits, and if it holds it becomes a row in Corrections + changes under the name of whoever raised it. It does not end anyone’s support. An objection is a dispute with a finding that survives that check, or any request to change, soften, delay or withdraw one. Every supporter is told, in writing and before anything is accepted, that the first time it objects its support ends and its objection is published in full alongside the finding. Nothing is withdrawn, softened, or delayed. A supporter that will not agree to this in advance is not accepted. The line between the two is whether the claim can be checked against the files on disk — the same line the harness already draws — and where it is genuinely unclear it is treated as a correction, and the checking is published.
10. Pre-registration precedes the ask. The INTENT file for a quarter’s Calibrations — the models, the probe set, the questions — is frozen before any funding application is filed in that quarter. What gets measured is therefore never shaped, even unconsciously, by who is being asked for money.

Buying access is not being funded. Paying the going rate for an API endpoint is being a customer, and being a customer is what makes measurement possible at all. Ordinary paid access to a measured company’s models has never been vendor support and is not treated as such. The line falls at parity: access that is free, earlier than a paying customer’s, or on any term better than a paying customer’s, is a benefit from a measured party, is disclosed at any value including zero, and counts against the limits in condition 5.

Declining is published too. Support refused under these rules is recorded in the ledger with its date, the offering party, and the reason, exactly as accepted support is. A record of what was turned down is worth more than a record of what was taken, and it is the only evidence that rules bind when they are inconvenient.

When these conditions change. They are reviewed on Dec 03, 2026 and quarterly after that — on the calendar, and not at anyone’s prompting. An offer does not trigger a review, and no condition is revised while a specific offer is open. Any loosening proposed within thirty days of a particular offer is refused on that ground alone. Every change to this page is a row in Corrections + changes.

The request table

Everyone who will be asked, listed before anyone is asked

Condition 1 requires this list to exist before a single letter goes out, with the answers blank. Here it is. The list was fixed on Sep 09, 2026. Nothing has been asked of anyone on it. Every response column is empty because no request has been sent — not because none has been answered.

The rule for inclusion. Any company that makes, serves or routes a model this program has measured, or expects to measure under the current or next quarter’s pre-registered intent. That is the whole test. The list is not a judgment about which companies matter and it is not the industry — there are several hundred models in the catalogue this program draws from, and coverage widens over time. What matters is that membership follows a rule rather than a preference, so that a reader can check whether a company belongs on it and ask why one is missing.

Funders that neither make, serve nor route are not on this list: they create no measurement conflict, and they are named in the ledger below if and when they give. A company with no channel through which to ask — several publish weights with no support function at all — is recorded as such rather than left ambiguous.

29 companies as the list stands. It will grow, and growing is not a loophole: every addition is a row in Corrections + changes carrying its date and its reason, and a company added after the asking has begun is asked in the same words as everyone before it. A company is never removed from this table once it is on it — if it leaves scope the row stays and says so, because a list that can shed names is a list that can be tidied after the fact.

The letter

The words every company will receive

Condition 1 says everyone is asked in the same words. These are the words, published before the first letter is sent so that they cannot be adjusted for anyone afterwards. Each company receives exactly this text, subject line included, signed by Lawrence Ho. Where a company asks through a form of its own, the form is answered and this letter is attached unchanged. No letter has been sent.

Subject: Support request from Protocol STAMP, an independent AI evaluation program

Protocol STAMP is an independent program, published by Budday Budderson Studio LLC (New Mexico), that measures whether AI models verify their own work and report what they checked. We publish our method, our costs and our mistakes at https://protocolstamp.com/.

We are writing to every company that makes, serves or routes a model we measure, in these same words and in the same week. The full list was published on Sep 09, 2026, before anyone was asked: https://protocolstamp.com/independence/#request-table

We are asking for support in cash or in API credits on paid-tier or zero-data-retention endpoints. The terms are fixed and published, and they are the same for everyone:

1. Your support never pays for a measurement of your own models. Your models are always bought at retail with our own money; your credits are spent measuring other companies’ models, or on development work.
2. You receive a labeled line on our Independence and Support pages and in our site footer. Nothing else: no mention in any issue, no early access, no influence over what is measured.
3. Every issue states which route and whose credit paid for each measurement.
4. Telling us we made a factual error costs you nothing; we check it and publish the correction under your name. Objecting to a finding that survives that check, or asking for one to be changed, softened, delayed or withdrawn, ends the support, and the objection is published in full beside the finding. You agree to this in writing before anything is accepted.
5. No single company may provide more than 10% of our year’s resources, and all measured companies together no more than 25%.

The ten conditions in full: https://protocolstamp.com/independence/

Disclosure: this program uses Claude, made by Anthropic, as a tool to draft text, as our Colophon page describes. Anthropic is on the list above and is asked on the same terms.

If you would like to support the program on these terms, or to decline, a reply to this address is enough. Either answer is recorded, with its date, in the published table.

Lawrence Ho
Chief Human Investigator, Protocol STAMP
s@bud.day

Proposed, not in force

For the Dec 03, 2026 review: sponsorship of the weekly newsletter

The conditions above change only on the calendar, and the next review is Dec 03, 2026. This proposal is published now, before any offer exists, so that it is decided on that date with no offer open. Until then, nothing here is in force, and a sponsor gets only what the section “What a sponsor gets” says.

• Where. One labeled line per edition of the weekly email newsletter, if one is launched. Never in The Calibration, The Log or any issue, and never on an issue page.
• Who may sponsor. Only a company that neither makes, serves nor routes a model this program measures, is not on the request table, and is not owned by a company that is. A measured company may not sponsor the newsletter in any form, including through an agency.
• No reader offers from the measured. No sponsor line may carry a discount, code or credit for a model or service this program measures.
• The advertising rule. Advertising is otherwise placed by a third party without knowledge of content. Here the publisher would choose the sponsor, so the proposal is a stated exception for this one slot, and every sponsor is named in the ledger at $100 or more, under the objection clause.
• To be decided on Dec 03. Whether to adopt it, and a cap on newsletter sponsorship as a share of the year’s resources.

The ledger

Every supporter named, and everything declined

Supporters are named here with an amount band and a date, counted cumulatively across a calendar year, with support in kind valued at list price.

• Any company that makes, serves, or routes models this publication measures is named at any amount, including support given at no charge. There is no threshold, because the conflict is the relationship and not the sum.
• Any other business, foundation, or institution is named at $100.
• An individual is named at $100 with their consent, asked at the time of support. Without consent the entry reads “an individual supporter,” with the band and the date.

This replaces a flat $1,000 threshold, which would have let a company give $99 of credits — or free access worth nothing on paper — and appear nowhere.

Anonymity is available to individuals and to no one else. A company that makes, serves, or routes measured models may not support this work anonymously, in any amount or form; a business that asks to be unnamed is declined. Anonymous means the reader does not know, never that the publisher does not: every supporter’s identity is recorded, and support whose origin cannot be established is returned rather than listed. The ledger is revised within 30 days of any change, and the year’s totals appear in The Log’s cost section.

Current funders: the publisher.
Current support from companies this program measures: none.
Declined to date: none.

What a sponsor gets

Placement, and nothing else

A sponsor gets a labeled line on this page, on the support page, and in the home footer. Never a mention inside an issue, never a ranking, never a method, never early access. Nobody will be able to say you bought a result. Those terms are not concessions extracted from us; they are the whole value of the association — a name on an evaluation program is worth precisely as much as that program’s freedom to publish something the name would not have chosen, which is why the objection clause is written into every agreement before any money moves.

The ask

What support would fund, in the order it would be spent

What it costs, estimated. At full cadence — the bank fired on schedule, panels rather than a single grader, and several agent teams running at once — the program needs up to $15,000 a month in cash or API credit. That is an estimate, and it is labeled as one rather than printed as a finding: there is no charged figure yet, for the reason given in item 1 below. When a measured figure exists it will replace this one, and this estimate will stay on the page beside it, so anyone can see how far off it was.

1. A fully instrumented round, and then a price. 2,673 draws in measurement one carry no recorded cost, because the runner learned to record spend after they were taken — The Log prints that rather than folding it into a total that would look complete. The first thing money buys is a round instrumented end to end, after which this section stops describing work and starts quoting a figure.
2. The confidence instrument. Asking a model what it would stake on an answer, and scoring that against what actually happened. It is the difference between catching a compounding failure and reporting one afterwards.
3. The Gemini re-grade of both arms, and the judging panel on Luna. Vol 001, Finding 02: the arm comparison cannot be published as it stands, because the tier-1 grader sat 4.1 points below a three-judge panel — four times the size of the effect it was meant to reveal. In measurement one the grader cost multiples of the models it was grading. That is why the protocol question sits queued behind its instrument, and money is what un-queues it.
4. Publishing the run files. figures.json registers the SHA-256 of files nobody can yet download, so the digests cannot be recomputed off this site. That is stated as a real gap on the home page rather than implied away. Closing it is hosting and hours, not research.

Reader support and named sponsorship are both intended, on the terms above. This is a New Mexico LLC and not a charity: nothing here is tax-deductible, and anyone who tells you otherwise about a sponsorship is selling something. There is one address, nine characters long: s@bud.day.

People and machines

Who makes this

Findings are produced with AI assistance throughout, including on themselves: the graders and judges are AI models, and the analysis and drafting were produced with substantial AI assistance. A program that evaluates AI systems is itself largely built and operated by them, and a reader is entitled to weigh that. What is done about it: every number is computed from files on disk with the fingerprint of its sources; no figure may enter a document by being typed; the grader is validated by a panel rather than trusted; and errors in our own work are published rather than corrected quietly. The four names on the masthead and what each may do are on the About page.

Tools. The site and the harness were built primarily with Claude (Anthropic), bought at the public price and disclosed on the Colophon, which also explains the watermark Anthropic has embedded in its models’ text since Aug 02, 2026. Anthropic is a vendor this program measures or may measure. It has no sponsorship, funding, review, or endorsement relationship with this program, and using its tools does not exempt its models from any rule on this page, including recusal.